Skewness Kurtosis normality test

[Shafaq Malik]

Hi

I carried out sktest of normality and results are as follows

Skewness/Kurtosis tests for Normality
joint ------
Variable Obs Pr(Skewness) Pr(Kurtosis) adj chi2(2) Prob>chi2

pm 16,059 0.0000 0.0000 . .
roa 16,059 0.0000 0.0000 . .
roe 16,059 0.0000 0.0000 . .
rt 16,059 0.0000 0.0153 . 0.0000
rtm 16,059 0.0000 0.0000 . .
rta 16,059 0.0000 0.0000 . .
rts 16,059 0.0000 0.0000 . .
lev 16,059 0.0000 0.0000 . .
mv 16,059 0.0000 0.0000 . 0.0000
bmv 16,059 0.0000 0.0000 . .

I wonder why it's not showing values for chi2 and probability?

[Joseph Coveney]

I don't know, but what's going on with rt? You've got 16 059 observations and all you can get is a measly P = 0.015, unadjusted, for kurtosis.

[Nick Cox]

It may be that the implied values are too small to calculate.

Personal opinion: Such tests are practically useless for sample sizes like 16059. Almost any skewness and kurtosis that is slightly different from the normal reference values will produce overwhelmingly small P-values at that sample size. Significance at conventional levels can mean anything from your having slight nonnormality that isn't a problem to your being in Total Nightmare Territory.

You are better off thinking about measuring skewness and kurtosis directly and indeed looking at graphs of the data. Perhaps you're doing that any way, if so excellent.

[Bruce Weaver]

Hello Shafaq. Why do you want to test for normality? Thanks for clarifying.

[Shafaq Malik]

Dear Nick, graphs and skewness and kurtosis show that data is not normally distributed.

I would like to ask one more thing, I have read somewhere that normality check should be for residuals not for raw data, is that right?

Secondly, my data is panel data, how can I deal with my data for issue of non normal distribution without transforming it?

[Shafaq Malik]

Dear Bruce

I have to carry out panel data analysis and have been asked to conduct test for normality of data
Do you think that normality test is required in case of panel data or in case of Fama Macbeth estimates?

[Nick Cox]

It is generally true that conditional normality is of more interest than marginal normality. But sometimes marginal non-normality is part of the grounds for doing something different e.g . using a transformed scale and/or non-identity link function.

I am not an economist or econometrician and can't comment on Fama-Macbeth. Equally, the role of normality is discussed at length in every mainstream econometrics text I have ever glanced at.